DEC-0025: Prompt Claim Exclusivity and Exception-First Ops
Decision
The prompt relay now hard-enforces one active in-progress prompt per agent.
The internal prompt/ops surfaces should optimize first for exceptions requiring judgment, not only for aggregate reporting.
Context
The repo's governance language had already converged on a Kanban-like invariant:
one active prompt per agent, one context window per session. But that rule
lived primarily in policy and framework language rather than in the claim path
of scripts/prompt-queue.py.
At the same time, the internal dashboard and prompt views remained stronger at showing totals than at surfacing what the operator actually needs to decide:
- blocked/deferred/rejected work
- stale claims
- conflicting in-progress ownership
- queue-state drift
The human explicitly chose a tactical posture centered on "what needs judgment now?" and approved stronger enforcement.
Rationale
The cheapest place to prevent prompt thrash is the claim boundary.
If an agent can casually accumulate multiple active prompts, the system loses the very WIP discipline it claims to have. That degrades handoff fidelity, muddies ownership, and makes the queue less trustworthy as an external memory layer.
Likewise, a dashboard that mainly reports totals can feel healthy while hiding the exact state the operator most needs to resolve.
Consequences
Positive
- A previously policy-only invariant is now tool-enforced.
- The queue becomes a more trustworthy picture of actual active ownership.
- Internal ops views can now function more like an exception console than a passive archive browser.
Tradeoffs
- Agents must resolve active work before claiming new prompt work.
- Some metadata gaps remain visible without yet becoming hard validation failures. This is intentional while the corpus continues to normalize.
Implementation Standard
- The claim command should reject a new claim when the agent already owns
another
in-progressprompt. - Internal prompt views should surface stale claims, blocked/deferred/rejected work, and queue git drift prominently.
- Metadata confidence gaps may be shown as warnings before they become strict completion requirements.